fuse: allow requests during mount (needed for SELinux labels)

Change-Id: Ia1af402897e6a7290acf79617c34fdc804751729
BUG: 811217
Signed-off-by: Jeff Darcy <jdarcy@redhat.com>
Reviewed-on: http://review.gluster.com/3199
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Anand Avati <avati@redhat.com>

3 files changed, 136 insertions, 29 deletions

diff --git a/xlators/mount/fuse/src/fuse-bridge.c b/xlators/mount/fuse/src/fuse-bridge.c
index c9f7c89408d..c06da7eec58 100644
--- a/xlators/mount/fuse/src/fuse-bridge.c
+++ b/xlators/mount/fuse/src/fuse-bridge.c
@@ -17,6 +17,7 @@
    <http://www.gnu.org/licenses/>.
 */
 
+#include <sys/wait.h>
 #include "fuse-bridge.h"
 
 static int gf_fuse_conn_err_log;
@@ -3859,20 +3860,50 @@ unlock:
         return 0;
 }
 
+int
+fuse_get_mount_status (xlator_t *this)
+{
+        int             kid_status = -1;
+        pid_t           kid_pid = -1;
+        fuse_private_t *priv = this->private;
+        int             our_status = -1;
+
+        if (read(priv->status_pipe[0],&kid_status, sizeof(kid_status)) < 0) {
+                gf_log (this->name, GF_LOG_ERROR, "could not get mount status");
+                goto out;
+        }
+        gf_log (this->name, GF_LOG_DEBUG, "mount status is %d", kid_status);
+
+        if (read(priv->status_pipe[0],&kid_pid, sizeof(kid_pid)) < 0) {
+                gf_log (this->name, GF_LOG_ERROR, "could not get mount PID");
+                goto out;
+        }
+        gf_log (this->name, GF_LOG_DEBUG, "mount PID is %d", kid_pid);
+
+        (void)waitpid(kid_pid,NULL,0);
+        our_status = kid_status;
+
+out:
+        close(priv->status_pipe[0]);
+        close(priv->status_pipe[1]);
+        return our_status;
+}
 
 static void *
 fuse_thread_proc (void *data)
 {
-        char           *mount_point = NULL;
-        xlator_t       *this = NULL;
-        fuse_private_t *priv = NULL;
-        ssize_t         res = 0;
-        struct iobuf   *iobuf = NULL;
-        fuse_in_header_t *finh;
-        struct iovec iov_in[2];
-        void *msg = NULL;
-        const size_t msg0_size = sizeof (*finh) + 128;
-        fuse_handler_t **fuse_ops = NULL;
+        char                     *mount_point = NULL;
+        xlator_t                 *this = NULL;
+        fuse_private_t           *priv = NULL;
+        ssize_t                   res = 0;
+        struct iobuf             *iobuf = NULL;
+        fuse_in_header_t         *finh;
+        struct iovec              iov_in[2];
+        void                     *msg = NULL;
+        const size_t              msg0_size = sizeof (*finh) + 128;
+        fuse_handler_t          **fuse_ops = NULL;
+        struct pollfd             pfd[2] = {{0,}};
+        gf_boolean_t              mount_finished = _gf_false;
 
         this = data;
         priv = this->private;
@@ -3889,6 +3920,41 @@ fuse_thread_proc (void *data)
                 /* THIS has to be reset here */
                 THIS = this;
 
+                if (!mount_finished) {
+                        memset(pfd,0,sizeof(pfd));
+                        pfd[0].fd = priv->status_pipe[0];
+                        pfd[0].events = POLLIN | POLLHUP | POLLERR;
+                        pfd[1].fd = priv->fd;
+                        pfd[1].events = POLLIN | POLLHUP | POLLERR;
+                        if (poll(pfd,2,-1) < 0) {
+                                gf_log (this->name, GF_LOG_ERROR,
+                                        "poll error %s", strerror(errno));
+                                break;
+                        }
+                        if (pfd[0].revents & POLLIN) {
+                                if (fuse_get_mount_status(this) != 0) {
+                                        break;
+                                }
+                                mount_finished = _gf_true;
+                        }
+                        else if (pfd[0].revents) {
+                                gf_log (this->name, GF_LOG_ERROR,
+                                        "mount pipe closed without status");
+                                break;
+                        }
+                        if (!pfd[1].revents) {
+                                continue;
+                        }
+                }
+
+                /*
+                 * We don't want to block on readv while we're still waiting
+                 * for mount status.  That means we only want to get here if
+                 * mount_status is true (meaning that our wait completed
+                 * already) or if we already called poll(2) on priv->fd to
+                 * make sure it's ready.
+                 */
+
                 if (priv->init_recvd)
                         fuse_graph_sync (this);
 
@@ -4010,8 +4076,11 @@ fuse_thread_proc (void *data)
                 GF_FREE (iov_in[0].iov_base);
         }
 
-        iobuf_unref (iobuf);
-        GF_FREE (iov_in[0].iov_base);
+        /*
+         * We could be in all sorts of states with respect to iobuf and iov_in
+         * by the time we get here, and it's just not worth untangling them if
+         * we're about to kill ourselves anyway.
+         */
 
         if (dict_get (this->options, ZR_MOUNTPOINT_OPT))
                 mount_point = data_to_str (dict_get (this->options,
@@ -4019,11 +4088,10 @@ fuse_thread_proc (void *data)
         if (mount_point) {
                 gf_log (this->name, GF_LOG_INFO,
                         "unmounting %s", mount_point);
-                dict_del (this->options, ZR_MOUNTPOINT_OPT);
         }
 
+        /* Kill the whole process, not just this thread. */
         kill (getpid(), SIGTERM);
-
         return NULL;
 }
 
@@ -4492,8 +4560,15 @@ init (xlator_t *this_xl)
         if (!mnt_args)
                 goto cleanup_exit;
 
+        if (pipe(priv->status_pipe) < 0) {
+                gf_log (this_xl->name, GF_LOG_ERROR,
+                        "could not create pipe to separate mount process");
+                goto cleanup_exit;
+        }
+
         priv->fd = gf_fuse_mount (priv->mount_point, fsname, mnt_args,
-                                  sync_mtab ? &ctx->mtab_pid : NULL);
+                                  sync_mtab ? &ctx->mtab_pid : NULL,
+                                  priv->status_pipe[1]);
         if (priv->fd == -1)
                 goto cleanup_exit;
 
diff --git a/xlators/mount/fuse/src/fuse-bridge.h b/xlators/mount/fuse/src/fuse-bridge.h
index 72aa6be48f3..d20413055d8 100644
--- a/xlators/mount/fuse/src/fuse-bridge.h
+++ b/xlators/mount/fuse/src/fuse-bridge.h
@@ -66,7 +66,7 @@
 
 #define MAX_FUSE_PROC_DELAY 1
 
-#define DISABLE_SELINUX 1
+//#define DISABLE_SELINUX 1
 
 typedef struct fuse_in_header fuse_in_header_t;
 typedef void (fuse_handler_t) (xlator_t *this, fuse_in_header_t *finh,
@@ -116,6 +116,9 @@ struct fuse_private {
         int                  revchan_in;
         int                  revchan_out;
         gf_boolean_t         reverse_fuse_thread_started;
+
+        /* For communicating with separate mount thread. */
+        int                  status_pipe[2];
 };
 typedef struct fuse_private fuse_private_t;
 
diff --git a/xlators/mount/fuse/utils/mount.glusterfs.in b/xlators/mount/fuse/utils/mount.glusterfs.in
index 073c3609abf..aef0939db62 100755
--- a/xlators/mount/fuse/utils/mount.glusterfs.in
+++ b/xlators/mount/fuse/utils/mount.glusterfs.in
@@ -34,6 +34,36 @@ _init ()
     UPDATEDBCONF=/etc/updatedb.conf
 }
 
+# Mount happens asynchronously, so the command status alone will never be
+# sufficient.  Instead, we have to wait for multiple events representing
+# different possible outcomes.
+wait_for ()
+{
+    local daemon_pid=$1
+    local mount_point=$2
+
+    waited=0
+    while true; do
+        kill -s 0 $daemon_pid
+        if [ $? != 0 ]; then
+                echo "Gluster client daemon exited unexpectedly."
+                umount $mount_point &> /dev/null
+                exit 1
+        fi
+        inode=$(stat -c %i $mount_point 2>/dev/null);
+        if [ "$inode" = "1" ]; then
+            break
+        fi
+        if [ $waited -ge 10 ]; then
+            break
+        fi
+        sleep 1
+        waited=$((waited+1))
+    done
+
+    echo "$inode"
+}
+
 start_glusterfs ()
 {
     if [ -n "$log_level_str" ]; then
@@ -121,34 +151,33 @@ start_glusterfs ()
         cmd_line=$(echo "$cmd_line --volfile=$volfile_loc");
     fi
 
-    cmd_line=$(echo "$cmd_line $mount_point");
-    err=0;
-    $cmd_line;
-
-
-    inode=$(stat -c %i $mount_point 2>/dev/null);
+    cmd_line=$(echo "$cmd_line $mount_point")
+    err=0
+    $cmd_line
+    daemon_pid=$$
 
+    # Wait for the inode to change *or* for the daemon to exit (indicating a
+    # problem with the mount).
+    inode=$(wait_for $daemon_pid $mount_point)
     # this is required if the stat returns error
     if [ -z "$inode" ]; then
-        inode="0";
+        inode="0"
     fi
 
     # retry the failover
-    # if [ $? != "0" ]; then # <--- TODO: Once glusterfs returns proper error code, change it.
     if [ $inode -ne 1 ]; then
-        err=1;
         if [ -n "$cmd_line1" ]; then
             cmd_line1=$(echo "$cmd_line1 $mount_point");
-            $cmd_line1;
-            err=0;
+            $cmd_line1
+            daemon_pid=$$
 
-            inode=$(stat -c %i $mount_point 2>/dev/null);
+            inode=$(wait_for $daemon_pid $mount_point)
             # this is required if the stat returns error
             if [ -z "$inode" ]; then
                 inode="0";
             fi
             if [ $inode -ne 1 ]; then
-                err=1;
+                err=1
             fi
         fi
     fi